研究

未来三年的工作，源头都在这里。

系统背后的工作：论文、项目、用到的方法，以及它们最初要回答的问题。

• 从数字病理图像预测基因表达（分子生物学）

  Gene-expression assays are slow, expensive and logistically painful. Whole-slide pathology images are fast and cheap. This thesis asks whether the second can stand in for the first — recovering molecular-biology signal directly from gigapixel pathology images, under a hard constraint on compute and latency. The approach combines graph neural networks, vision foundation models, large-language-model techniques adapted for this task, and statistical/probabilistic methods.

• 模拟人类视觉认知，并区分真实知觉与幻觉

  A year-long study using generative models — autoencoders and GANs — to model how the human brain constructs visual understanding, and to distinguish genuine perception from hallucination.